Energy Singularity’s Yang Zhao on Taming Fusion: A 3-Hour Interview
Summary
- Yang Zhao’s core judgment is that controlled nuclear fusion has moved from “can it be done scientifically?” to “can we bring down the cost per kilowatt-hour?” ITER is pursuing Q≥10 with low-temperature superconductors, after €25B in investment and roughly 30 years of construction; with high-temperature superconductors, doubling the magnetic field could reduce linear dimensions to about 30%, and volume and construction costs to roughly 2%. “Once you get the cost per kilowatt-hour down to the level of coal-fired power, you have commercialization.”（哪天把度电成本降到跟火电一样了，你就商业化了。）
- Energy Singularity has broken a decade-plus wager into financeable, verifiable milestones through 3 machines. Honghuang 70 cost about RMB120M in actual investment and has validated the engineering feasibility of a fully high-temperature-superconducting tokamak; the company plans to build the Q≥10 Honghuang 170 for about RMB3B by the end of 2027, then deliver the roughly 500MW Dongfang 380 demonstration power station in 2030–2035. “The third machine is a product to us.”
- Honghuang 70 and the 21.7T Jinghong magnet prove the ability to build, not that fusion has achieved energy break-even. Honghuang 70’s parameter is roughly 10^17, still four orders of magnitude below 10^21; Jinghong surpassed CFS/MIT’s previous roughly 20T record while also targeting high engineering current density, validating the design and manufacturing process for Honghuang 170’s 18 magnets with a peak toroidal field of about 23T. “The first complete ship made of steel” is Yang Zhao’s most accurate description of Honghuang 70.
- Honghuang 170 deliberately avoids new scientific risk, instead breaking whole-machine delivery into a series of high-parameter engineering milestones. It adopts a conservative physics design supported by extensive experiments on ITER and other machines, then validates engineering parameters one by one through prototypes for magnets, heating and power systems; these 3 systems account for roughly 90% of hardware costs. Yang calls fusion commercialization “an engineering problem with a solution,” with the real risks being insufficient people or money, systems that cannot be built, or a machine that fails to hit design specifications.
- Fusion startups exhibit a powerful first-mover feedback loop, and the same route is unlikely to accommodate a second player for long. CFS in the US and Energy Singularity are both betting on high-temperature-superconducting tokamaks; Helion’s field-reversed configuration route, based on the public data cited by Yang, has reached roughly 10^17 and may not yet have reached 10^18, leaving 4 orders of magnitude of scientific risk if extrapolated to 10^21, so he “doesn’t fully understand” how its 2028 power-generation target will be achieved. Talent, knowledge and capital will concentrate around the first mover: “The problem the second player is trying to solve may already have been worked on by the first player for 2 years.”
- From the first kilowatt-hour to a scalable power station, fusion must clear 3 hurdles: high Q, long-duration steady-state operation and fuel. Q=1 is only break-even at the plasma level; power stations typically require Q=5–30. Operation must also advance from about 1,000 seconds to 10,000 seconds, 100,000 seconds and eventually week- or month-long runs. Energy Singularity ultimately hopes to use deuterium-deuterium fusion, roughly an order of magnitude harder than deuterium-tritium, to eliminate tritium regulation, tritium-breeding plants and related costs. “There is an enormous amount of deuterium in seawater.”
- AI and fusion form a two-way infrastructure chain: AI is pushing electricity toward a bottleneck, while AI itself can lower the cost of fusion R&D. Yang believes exponentially growing compute will eventually collide with energy supply, while cheap energy, once available, will inevitably be consumed by new applications; conversely, AI can support real-time plasma control, replace expensive diagnostic hardware and compress hundreds of physical experiments into a small number of real-world trials. “As long as you can provide the energy, it will all be used up.”
Deep dive
1. Energy Singularity Started in a “Move-In Ready” Lingang Factory
Energy Singularity was founded in June 2021 and has completed 2 funding rounds. At the time of the interview, it was valued at RMB2.2B to just under RMB2.3B and was still in the R&D investment phase. Yang Zhao was 34 and said his MBTI “should be INTJ.”
The company chose Shanghai’s Lingang primarily because of schedule constraints, not industrial branding: to build and operate its first machine within 2 years, factory space, power and infrastructure had to be immediately available, leaving few viable locations.
The team grew from 4 founders to about 100 people when Honghuang 70 was completed and roughly 150 at the time of the interview. Only about 20 people focus directly on physics; most are engineers working on structures, cryogenics, vacuum, magnet manufacturing and related fields.
2. Controlled Nuclear Fusion Is First a System for Releasing Nuclear Energy on Demand
Yang Zhao explains fission and fusion through mass-energy equivalence: the total mass of the reaction products is lower than that of the reactants, and the difference becomes kinetic energy through E=mc². Because the square of the speed of light is so large, a tiny loss of mass releases enormous energy.
Fission splits heavier atomic nuclei, while fusion collides and combines 2 lighter nuclei. The former underlies atomic bombs and fission power plants; the latter first proved that energy could be released through the hydrogen bomb, before the focus shifted to controlled output.
“Controlled” does not mean making the reaction gentler. It means producing a designed, stable output of 500MW or 1GW. Without control, nuclear energy is released all at once in an extremely short period—that is a weapon.
3. Earth Cannot Replicate Stellar Gravity, Leaving Magnetic or Inertial Confinement
Stars use gravity to compress vast quantities of atoms to sufficiently high density. Earth lacks that condition, so humans mainly use inertial confinement to create extremely high temperatures and pressures for an instant, or magnetic fields to suspend hot plasma away from solid equipment.
The differences among magnetic-confinement routes are fundamentally differences in magnetic-field geometry. Tokamaks have achieved the highest triple product in existing experiments, leading other magnetic-confinement routes by roughly 2, or even 4, orders of magnitude.
Yang therefore sees magnetic confinement as the main civilian power-generation route: it has more scientific evidence, more than 100 machines, and is the only route to have genuinely approached energy break-even.
4. Tokamaks Use a “Doughnut” Magnetic Field to Confine Plasma in a Ring
Originating in the Soviet Union, the tokamak uses a ring-shaped magnetic field resembling a closed “doughnut.” Hot plasma collides continuously inside the ring, generating fusion energy without directly striking the vacuum-chamber wall.
It has not only a toroidal magnetic field along the doughnut, but also a critical toroidal current. Only the combination of toroidal field, toroidal current and 2D axisymmetric structure constitutes the tokamak configuration Yang describes.
Magnets are therefore the most critical system in the machine. Change the magnet material, and the operating temperature, vacuum, structure, heat dissipation and interfaces with other subsystems all need to be redesigned.
5. Tokamaks Have Passed Through 3 Generations of Materials: Copper, Low-Temperature Superconductors and High-Temperature Superconductors
Early machines used copper because it was simple to engineer and suitable for quickly obtaining plasma-experiment data. But high current heats copper, so at high parameters it can typically run for only seconds to more than 10 seconds; continued operation could even melt the magnets.
Low-temperature superconductors solved the continuous-heating problem, but required high vacuum, thermal shields and an environment around 4K, or -269°C. EAST began construction around 2002 and was completed in 2006–2007 as the first fully low-temperature-superconducting tokamak, followed by KSTAR and JT-60SA.
The second-generation high-temperature-superconducting material yttrium barium copper oxide reached industrial production at the 100-kilometer scale after roughly 2015, especially around 2018. Its actual operating temperature remains very low, but its critical field can be at least twice that of low-temperature superconductors.
6. ITER Proved That “Anything Is Possible at Any Cost,” While Exposing the Economic Limit of Low-Temperature Superconductors
ITER is led by the EU with participation from China, the US and 4 other parties. Its design was completed in the 1990s, construction began in 2006, its latest schedule has slipped to around 2034, and cumulative investment has reached about €25B.
Yang calls it the second-most-expensive international joint project after the International Space Station. Its Q≥10 engineering target is important, but a construction cycle of roughly 30 years also means new technologies could be delivered before it is completed.
The critical field of low-temperature superconductors is limited, so achieving sufficient energy gain requires making the machine extremely large. Yang estimates that if such a research machine were used directly for power generation, its cost per kilowatt-hour could exceed coal-fired power by 100x—a cost gap spanning 2 orders of magnitude.
By contrast, MIT and CFS’s SPARC seeks similar performance using high-temperature superconductors, with a budget of roughly $1B. Energy Singularity’s own estimate is lower, at about $400M for a comparable target.
7. The Triple Product Determines Q; High Magnetic Fields Are the Fastest Lever for Shrinking the Machine
The triple product governing fusion gain is the product of plasma density, temperature and confinement time. With deuterium-tritium fuel, reaching roughly 10^21 brings Q close to 1; moving toward 10^22 could make Q rise rapidly, like an avalanche.
The host’s question was critical: if the 3 parameters are multiplied together, why not simply raise one of them? Yang explains that pushing any single parameter higher often drags down the other 2, so engineering must optimize the product as a whole.
Because no first-principles model can directly derive the performance of an entire machine, the industry relies on decades of data from more than 100 machines and more than 1,000 experiments to fit scaling laws. For tokamaks, the triple product is approximately proportional to magnetic field to the 3.5th power and linear size to the 2.5th power.
Doubling the magnetic field can reduce linear dimensions to about 30% and volume to roughly 2%. Since the cost of non-standard equipment broadly tracks mass and volume, this is the physical basis for high-temperature superconductors potentially cutting costs by 2 orders of magnitude.
8. Q=1 Is Only Physical Break-Even; Power Stations Need Much More
Q is the ratio of fusion output power to input power. Q=1 means break-even at the plasma level, but excludes electricity conversion and whole-plant losses, so the overall facility may still consume net power.
Yang gives a power-station design range of Q=5–30. The industry generally treats Q=10 as a representative threshold for engineering feasibility: “Above 1 is a basic point,” but actual power generation requires far more than 1.
One of the highest public results in magnetic confinement was JT-60U’s deuterium-deuterium experiment, which inferred a deuterium-tritium-equivalent Q of 1.25. In inertial confinement, the US NIF achieved approximately Q=1.5 using an actual deuterium-tritium target.
JT-60U, TFTR, DIII-D and the UK’s JET had already approached 10^21. Those results in the 1990s gave the international community confidence to propose the Q>10 ITER.
9. Inertial Confinement Holds the Q Record but Is Not Yang Zhao’s Civilian Route
NIF and China’s counterpart “Shenguang” machines use large numbers of lasers to compress a deuterium-tritium target synchronously for an extremely short period. Each reaction lasts about a nanosecond, or 10^-9 seconds, making the approach better suited to studying extreme conditions and advanced nuclear weapons.
NIF’s Q uses the laser energy delivered to the target as input and excludes the losses involved in converting electricity into laser energy. Yang puts that conversion efficiency below 3%, making the end-to-end energy balance far less attractive than Q=1.5 suggests.
Inertial confinement can study how to raise the burn fraction from the roughly 1%–2% of early nuclear weapons to 80% or higher. But it is difficult to operate steadily and constrained by laser efficiency, so it is not the team’s power-generation route.
10. Six Decades of Experimental Consensus Have Made the Tokamak the Main Track
The Soviet Union proposed the tokamak in the 1960s, and early machines significantly outperformed contemporary magnetic-confinement routes. Better experimental results attracted more funding, machines and talent, which further improved performance.
More than 100 tokamaks have been built globally. The Institute of Plasma Physics in China, nuclear-industry research institutes, and high-parameter machines in the US, UK and Japan have all advanced along this route, making ITER a tokamak as well.
Yang does not equate consensus with the final answer, but treats it as an asset that lowers scientific risk. Experiments that have genuinely reached near 10^21 reduce the unknown physics involved in extrapolating directly from low parameters to a commercial machine.
11. Honghuang 70 Proved for the First Time That a Fully High-Temperature-Superconducting Tokamak Could Be Built
Energy Singularity completed Honghuang 70 in February–March 2024 and achieved first plasma in June. “70” refers to its major radius of roughly 70 centimeters, while “Honghuang” describes a state of abundant but chaotic energy that must be tamed.
Yang’s analogy is that all ships used to be made of wood, and now the industry wants to build an aircraft carrier out of steel. Honghuang 70 was not an aircraft carrier but “the first complete ship made of steel,” proving that the new material could be processed, assembled, launched, operated and brought back.
When asked whether it had already validated smaller volume and lower cost, Yang explicitly rejected that conclusion. The comparison must be made at equal performance; Honghuang 70 validated engineering principles, while the commercial proof of cost and Q belongs to Honghuang 170.
12. The 2021 Startup Thesis Was Not New Physics but a 2-Order-of-Magnitude Cost Reduction
Yang’s judgment at the time was that humanity already had a solid scientific and engineering foundation to build fusion “at any cost.” What was missing was commercial viability: bringing the cost per kilowatt-hour down to the level of coal-fired power or below.
Gradual cost reduction would struggle to bridge a 100x gap. High-temperature superconductors, however, could shrink the machine and its cost in one move through a material change, then deliver another order-of-magnitude reduction through scaling. That was the “engineering inflection point” he described.
Once the question became how to commercialize in the shortest time at the lowest cost, Yang believed a startup was better suited than a university or research institute. He compared this organizational division of labor to SpaceX’s push to industrialize spaceflight.
The company’s value boundary was therefore broad: any design, material, process or system that can continuously improve fusion’s cost-performance ratio and reduce the cost per kilowatt-hour falls within Energy Singularity’s mandate.
13. In-House Development Is a Cost-Model Requirement, Not a Technical Obsession
Energy Singularity insists on developing its own whole-machine design, magnet design and manufacturing, core subsystems, commissioning and experimental operations because even a minor change to one core system can alter the interfaces of others and materially change whole-machine cost.
Yang wants to push costs back toward raw materials: “All the knowledge and information was worked out by my own team.” Only then does the machine stop being a black box, and the team knows which system to adjust when raising parameters in the next generation.
External procurement is concentrated in highly competitive machining, welding, mature components and raw materials. Where suppliers are few, bargaining power is strong and the capability will be used repeatedly over the long term, the team prefers to control it internally.
14. Training in Theoretical Physics Provides a Judgment Framework, Not a Ready-Made Fusion Answer
Yang studied physics at Peking University after being admitted through physics competitions, then completed a PhD at Stanford in 2017. His research focused on quantum gravity, string theory and the intersection of quantum gravity and quantum information—far from engineering fusion.
He values physics’ approach of understanding a complex world with the fewest assumptions, while acknowledging the limits of pure reductionism: across multiple scales, new dominant properties emerge, and directly building an effective theory at that scale is often more accurate than forcing a derivation from the bottom up.
The host mentioned the connection between early deep learning and physics prizes. Yang believes Boltzmann machines do have physical analogies and analytical tools, but AI has mainly been advanced by computer scientists and AI researchers and should not therefore be treated simply as a branch of physics.
Fusion is “mostly a physics problem,” spanning nuclear physics, plasma physics, superconductivity, electromagnetism and materials. The team does not invent every material; it selects existing materials and solves the problems of structure, mechanical loads and heat dissipation.
15. A Startup That Wasn’t “Unique Enough” Made Him Redefine the Founder’s Commitment
After graduating from Stanford, Yang initially considered starting a company in the US. He returned to China at the end of 2018 and, with support from GSR Ventures, founded a company combining AI and music education, going through roughly 3 years of rapid change in online education.
He later concluded that a major startup cannot be merely something one “can do” or is “interested in.” It must remain the only choice one refuses to abandon even when faced with difficulties and other options. Good times are not the test; the real questions emerge in adversity.
In 2021, he spent about 6 months turning fusion from a “sometime in my lifetime” idea into a checklist: If it cannot be done now, what exactly is missing? If the paths for talent, technology, materials and capital are all viable, “why not start doing it now?”
16. Early Research Found No No-Go, but No Ready-Made Talent Pool Either
The early team visited research institutes and suppliers to determine whether China would be “choked” by shortages of talent, technology or raw materials. The conclusion was that raw materials and core suppliers were broadly available, while knowledge about the new route was largely contained in public academic work.
Talent was not readily available, but neither was it an insurmountable constraint. Fusion commercialization depends more on engineering execution, and China has an advantage in its supply of engineers; the team could train people with strong fundamentals internally.
This is not an industry won through a secret patent, but through “hundreds or thousands of Edison-style problems.” Knowledge accumulates rapidly with every machine, and scarce talent and capital create powerful positive feedback for the first mover.
17. The Industrial Chain Has Yet to Form, So Startups Must Deal Directly with Raw Materials
Around 2015, fusion research in China was led mainly by the Institute of Plasma Physics in Hefei and nuclear-industry research institutes in Chengdu, with scientific research still the primary goal. Yang calls Energy Singularity China’s first startup focused on commercial nuclear fusion.
Traditional research machines often assign an entire subsystem to another research institute. Energy Singularity instead buys high-temperature-superconducting tapes, steel, resin and electronic components directly, then sends drawings to competitive machining and welding factories. A mature industrial chain has yet to emerge.
18. A Tokamak Must Drill Continuously from Physics Targets Down to More Than 30 Subsystems
The first layer is physics design: define the machine’s core experimental objective, then translate it into physical parameters such as temperature, density, magnetic field and plasma shape.
The second layer is conceptual design: determine the parameters for roughly 10 primary systems and more than 30 secondary systems, divide responsibilities among the vacuum chamber, magnets, cryogenics, fueling, diagnostics and controls, and design interfaces so they do not “fight each other.”
The third layer is engineering design: turn conceptual requirements for flow, temperature and velocity into valve boxes, liquid-helium tanks, refrigerators, structural components and drawings, then proceed to manufacturing, subsystem acceptance, final assembly, integrated commissioning and experimental operation.
For areas that are difficult to calculate accurately, the team first builds small samples, compresses simulation-experiment error to roughly 10%–20%, validates the model and then extrapolates to full-scale components. Design, simulation and process experiments repeat in a continuous loop.
19. Honghuang 70 Used Concurrent Engineering to Compress a Traditional Serial Schedule into 2 Years
The project entered design around March 2022, and most engineering design was complete by the end of 2022 or early 2023. But the team did not wait for “design freeze”: systems that matured early began processing in the second half of 2022.
Manufacturing continued through the end of 2023. Final assembly ran from September 2023 to February 2024, followed by integrated commissioning and experiments from March to June, culminating in first plasma.
The team numbered only about 40–50 at the design peak and roughly 100 when the machine was completed. Yang said bluntly that if they had waited for all design to finish before machining and for all manufacturing to finish before assembly, “it definitely could not have been done in 2 years.”
20. The Most Expensive Failures Often Surface Only at the Final Step
Yang describes the engineering rhythm this way: “The closer you get to the physical state, the bigger your problems become, and the more problems you have.” Design changes are cheapest during the design phase; once physical equipment is delivered, room for modification is smallest while rework pressure is highest.
The first full-scale toroidal-field magnet had passed every prior quality inspection, but its performance deteriorated sharply after the final process. The team did not install it in the machine, instead retaining it as a test article and manufacturing an additional production unit.
The cause has not been identified with 100% certainty. The team suspected mechanical damage from machining or transport vibration, so it added vibration protection across the full chain of internal manufacturing, external suppliers and transportation. Later magnets did not show the same degradation.
For Yang, the most torturous part is not that a problem is difficult, but “not knowing what went wrong.” Once the mechanism is correctly hypothesized and a verifiable solution is found, the rest is a matter of spending the time to execute it.
21. First Plasma Is Whole-Machine Acceptance, Not a Q Breakthrough
The neutral gas injected into the vacuum chamber is initially invisible. An electron gun or microwave first strips electrons from some atoms, after which the magnets and other systems operate as designed, eventually forming visible, glowing plasma close to the intended magnetic configuration.
First plasma means the vacuum, magnets, power supplies, pre-ionization, diagnostics and controls can work together to meet startup conditions. It is the moment when a complete machine changes from static equipment into a tokamak.
Honghuang 70’s discharge at the time lasted less than 1 second, with a parameter of roughly 10^17—4 orders of magnitude below 10^21—so it had nothing to do with a meaningful Q value. One subsequent goal was to push it toward steady-state operation at the 1,000-second scale.
22. Honghuang 70’s Greater Asset Was a Team That Had Completed the Full Loop
High-temperature-superconducting materials are fragile and difficult to process, while also having to work under low temperature, strong magnetic fields, mechanical loads and thermal stress. Honghuang 70 proved that the material could form a complete device and start stably, not merely produce an individual coil.
The machine’s physics and engineering designs were developed by the team, core systems were processed internally, assembly was deeply managed and experiments were fully led by the team. During the upgrade, even assembly was completed in-house, so the next machine would no longer be a system-level black box.
The machine’s localization rate exceeded 96%. Imports consisted mainly of cryogenic pumps, cryogenic valve boxes and a small number of sensors. Domestic substitutes exist, but imported components remain more stable and better at maintaining parameters over long periods in critical positions.
Yang believes that at the time, Energy Singularity was the only complete team in the world that had designed, manufactured and operated a fully high-temperature-superconducting tokamak from scratch.
23. A 3-Step Roadmap Separates the Principle Prototype, Gain Validation and Commercial Power Station
Honghuang 70 is a principle prototype whose task was to validate whether high-temperature superconductors could form a complete tokamak. It completed its first-stage delivery in 2024.
Honghuang 170 is intended to achieve Q≥10 at the lowest possible cost and was planned, as of the interview, to be completed in about 3 years, by the end of 2027. If the result matches the design, its core technology and cost would be sufficient to support a demonstration power station.
Dongfang 380 is planned for delivery in 2030–2035, with electrical output of about 500MW, equivalent to a medium-sized coal-fired power station. It is not an experimental machine for continued fundraising demonstrations, but equipment to be sold to nuclear-power or other energy owners.
If the final route uses deuterium-deuterium fusion without tritium, potential owners could also expand to local energy SOEs and other state-owned enterprises. Energy Singularity’s role is to supply the tokamak core equipment, not necessarily to own the power station outright.
24. The 21.7T Jinghong Magnet Pushed High-Field Capability Past CFS’s Record
Jinghong reached 21.7T, exceeding the roughly 20T achieved by CFS and MIT with TFMC at the end of 2021. Yang said that at the time, only Energy Singularity and CFS had teams capable of manufacturing magnets with large apertures above 20T.
A “large aperture” means leaving meter-scale space in the middle for the vacuum chamber and plasma, rather than the millimeter- or centimeter-scale apertures common in materials experiments. As the aperture expands, the difficulty of handling asymmetric forces, heat dissipation and machining rises sharply.
Honghuang 170 will comprise 18 toroidal-field magnets, with a peak whole-machine magnetic field of about 23T. Jinghong will not be installed directly in the machine, but its near-target single-unit parameters prove that the same design method and manufacturing process can deliver production magnets.
25. High-Field Magnets Must Push Force, Heat and Current Density to the Edge Simultaneously
High magnetic fields require high current, while electromagnetic force rises with both field and current. Adding more steel can increase load-bearing capacity, but it occupies cross-sectional area that cannot carry current, lowering engineering current density and forcing the machine to become large again.
Jinghong was designed not only for maximum magnetic field, but also for very high engineering current density. At the same total current, higher current density means a smaller conductor cross-section, which is what allows the tokamak to truly shrink.
At a single-strand current above 20kA, resistance of roughly 1 nano-ohm can produce close to 1W of heat. To hold total heat generation to the 100W range, resistance from all joints and machining must be reduced to about 100 nano-ohms, with helium flow channels arranged to carry away heat.
Yang compares the process to “baking bread”: when metals are fused, oven temperature, heating curve, holding time and cooldown steps all affect performance. Too high a temperature damages the superconducting tape; too low a temperature prevents proper fusion.
26. Honghuang 170 Converts Scientific Risk into Engineering Risk That Can Be Accepted Step by Step
The physics design follows a conservative path similar to ITER’s 30-year-old design, supported by extensive experiments. The team does not want to bet on new physics simultaneously; it only requires engineering systems to accurately deliver the input parameters required by the model.
The first risk is whether individual subsystems can reach extremely high parameters. Each critical system is therefore first built as a prototype, and prototype specifications can even exceed those of the production machine. Jinghong is the pre-research validation for the toroidal-field system.
The second risk is whether systems can be assembled without losing performance. Honghuang 70 has already provided experience with interface conflicts, installation sequence, inspection windows and trade-offs, reducing whole-machine complexity from something completely unknown to a loop the team has completed once.
Magnets, heating and power systems account for about 90% of Honghuang 170’s hardware cost, making the first prototypes of the 3 categories especially important. Other systems, including cryogenics and diagnostics, will be validated according to the same logic.
27. Without Honghuang 70, 4 People Could Not Raise RMB3B—or Justifiably Bet Directly on Honghuang 170
Honghuang 70 had an original budget of about RMB150M and ultimately cost about RMB120M. Honghuang 170 requires roughly RMB3B. A 4-person team that had never built a tokamak would have struggled to raise that amount and had no basis for believing it could deliver world-class parameters in one shot.
The engineering gap between the 2 machines is visible in their magnetic fields: Honghuang 70 peaked at about 3.1T, while Honghuang 170 targets about 23T. The credible financing chain is to prove whole-machine capability first, then fill in high-parameter capability through prototypes such as Jinghong.
The RMB3B will not be raised in a single round. At the time of the interview, the team was pursuing a single round of roughly RMB1B–RMB1.5B. Yang acknowledged that the first commercial product would still be 2030–2035 away, demanding substantial patience from investors and a long exit horizon.
28. 3 Low-Temperature-Superconducting Machines Built Steady-State Experience but Did Not Solve Miniaturization
EAST, KSTAR and JT-60SA all support low-temperature-superconducting engineering validation and the accumulation of experience for ITER. EAST has pushed long-pulse operation beyond 1,000 seconds, also showing that superconducting magnets are indispensable for steady-state operation.
In Yang’s view, there is little theoretical disagreement over the advantages of high-temperature superconductors—higher magnetic fields, smaller volume and lower cost. The real lack of consensus is whether this fragile material can be processed, how quickly it can be made, and ultimately who can make it.
29. Dongfang 380 Is Priced as a Complete Power Station, Not an Experimental Machine
Compared with Honghuang 170, Dongfang 380’s linear dimensions will nearly double, its volume could expand by roughly 10x, and its peak magnetic field will rise from about 23T to about 29T.
Honghuang 170 only needs to reach experimental parameters for short periods, allowing it to omit long-duration water cooling and energy-extraction systems. Dongfang 380 must be designed for continuous power generation, incorporating steady-state cooling, heat removal and full engineering support.
The team hopes to sell it for RMB40,000–RMB50,000 per kilowatt. A 500MW machine would cost about RMB20B–RMB25B, similar to the unit cost of China’s first fourth-generation high-temperature gas-cooled fission demonstration reactor. Estimated manufacturing cost is RMB10B–RMB20B.
Dongfang 380 will be custom-built after securing an owner’s order, rather than being built independently by Energy Singularity and marketed afterward. The company’s own core upfront capital requirement remains the roughly RMB3B needed for Honghuang 170.
30. CFS Is the Closest Peer; Helion Carries Greater Scientific Extrapolation Risk
CFS and Energy Singularity have both chosen high-temperature-superconducting tokamaks. SPARC began construction in 2022, had an external target of completion around 2026 and a stated budget of roughly $1B, making it the most comparable international project to Honghuang 170.
Helion uses a linear field-reversed configuration, or FRC. Yang cites the highest parameters in public academic FRC materials at roughly 10^17, possibly not yet reaching 10^18; that remains about 4 orders of magnitude below the 10^21 vicinity of deuterium-tritium break-even.
His analogy is designing an aircraft for 10,000 meters using only flight data from 0–10 meters: one might completely miss new factors such as thinning air and falling temperatures. Extrapolating from 10^17 to 10^21 could likewise encounter emergent physics absent from the original equations.
Because Helion has published few papers and technical materials, Yang did not assert that it will necessarily fail. But he admitted that he “doesn’t fully understand” how it will achieve energy break-even and considers its goal of building a fusion power station in 2028 “extremely aggressive.”
31. China and the US May Commercialize First, but Will Form 2 Relatively Independent Markets
Western commercial fusion has raised nearly $6B cumulatively across about 40 startups. China has fewer than 10 startups, with total funding of roughly RMB10B, leaving a significant gap in capital scale.
China is unlikely to rely on the US to build its fusion power stations, and the US is similarly unlikely to import an entire set of core technologies from China. Both have demand, capital, talent and supply chains, so each is likely to produce domestic winners.
The routes will not be completely separate: leading US companies often have counterparts in China. Energy Singularity corresponds to CFS in high-temperature-superconducting tokamaks, while Chinese teams are also researching FRC routes similar to Helion.
32. Cost Advantages Come from Design Constraints and the BOM, Not a Slogan About Chinese Manufacturing
Honghuang 170’s objective is to become the world’s lowest-cost, highest-value Q≥10 machine using about RMB3B. Compared with CFS’s publicly stated budget of roughly $1B, the team believes its own design is likely to be more economical.
This judgment is based on the granular BOM developed through Honghuang 70. Overall design, material selection, suppliers, processes and experimental operations were all optimized around minimum cost from the outset, rather than relying on procurement savings after completion.
Competitive steps are outsourced, while core capabilities with monopolistic bargaining power and repeated long-term use are developed in-house. Yang says the current design is already close to their perceived lower bound: “It is relatively difficult to optimize further.”
Commercial demand is unusually simple: the demand is real and the product is clearly defined. Energy “will definitely be used up as long as it can be supplied.” If the machine is delivered to specification and its cost per kilowatt-hour is below coal-fired power, it has clear buyers; failure risk is concentrated in resources and engineering execution.
33. AI’s Exponential Growth Will Ultimately Hit Energy Supply, Not Demand
Yang believes that if chip, data and compute capacity continue expanding, AI’s more distant hard bottleneck will become energy. If data-center electricity consumption rises from a few percent to tens of percent, existing infrastructure will struggle to support it.
Whether energy supply can expand ultimately depends on cost. If cost does not fall, the same investment cannot buy orders of magnitude more electricity; only a substantial reduction in the cost per kilowatt-hour can enable supply to scale sharply.
He compares energy to computing power: every performance increase quickly brings applications that consume the new capacity. “As long as you can provide the energy, it will all be used up”; there is no such thing as excess supply.
Support from US internet and AI companies for new energy and fusion companies can also be viewed as early positioning for the next electricity bottleneck. Sam Altman’s $375M investment in Helion was the representative bet highlighted at the opening of the episode.
34. AI’s Value to Fusion Is Faster Control, Diagnostics and Experimental Iteration
Real-time plasma control requires extremely high speed, while traditional complex physics models are difficult to calculate in time. AI can compress complex processes into equivalent operators that infer quickly while maintaining high accuracy.
Yang mentions DeepMind’s work on a European tokamak: pure AI control achieved plasma configurations that previously required years of accumulated experience with very few experimental iterations, producing 2 papers in Nature.
AI could also replace expensive diagnostic hardware, as in image and medical enhancement, using algorithms to deliver diagnostic results with higher accuracy and resolution.
If a digital model of a specific machine were calibrated with a small number of real experiments, the 100 experiments previously required might be reduced to 2 physical trials plus extensive simulation. Energy Singularity planned to use AI to replace part of its control algorithms on its own machines, but had not yet partnered with an external AI company.
35. 3 Hurdles Remain Between Fusion and the First Kilowatt-Hour: High Q, Long Pulses and Deuterium-Deuterium Fuel
First, the machine must reach a sufficiently high Q and stably control the plasma. Honghuang 170 is responsible for validating Q≥10, the first engineering threshold at which output is far greater than input.
Second, a demonstration power station cannot rely on short pulses. The industry has a record of roughly 1,000 seconds, but must still advance to 10,000 seconds, 100,000 seconds and eventually week- or month-long operation. If any subsystem shuts down first, the entire machine cannot continue generating power.
Third, Energy Singularity hopes eventually to shift from deuterium-tritium to deuterium-deuterium. Tritium is tightly regulated because deuterium-tritium can be used to make a hydrogen bomb. It also cannot stably exist in nature, so a deuterium-tritium power station must bombard lithium-6 with neutrons to achieve a tritium breeding ratio above 1 and recover the tritium as fuel.
Deuterium-deuterium fusion is roughly an order of magnitude harder than deuterium-tritium, requiring the triple product to rise from about 10^21 to 10^22, but it could eliminate tritium plants and related safety costs. Yang says the deuterium in seawater is sufficient for humanity to use for more than 10 billion years.
36. Fusion’s “ChatGPT Moment” May Arrive Before Truly Cheap Power
Yang draws a comparison with fission history: after the first fusion demonstration power station is built, it may take roughly 10 years to produce a replicable, stable commercial power station. The detailed cost-reduction curve after Dongfang 380 is still unclear.
The first demonstration power station might generate electricity at RMB0.4–RMB0.5 per kilowatt-hour, still above coal-fired power but already near the inflection point. Once a scalable path becomes visible, energy owners may quickly begin requesting multiple units.
An earlier consensus milestone could be SPARC or Honghuang 170 achieving Q≥1 or even Q≥10 at a technical cost close to coal-fired power. It would be like ChatGPT: not AGI itself, but proof to the market that AGI is achievable.
If the cost per kilowatt-hour merely matches coal-fired power, fusion could replace part of coal generation. If it falls another order of magnitude, Yang believes the energy structure would be rewritten, enabling industrial synthesis of food, manufacturing methods previously uneconomic to calculate and interstellar activity.
37. After Power Generation, the Long-Term Vision Is Propellantless Propulsion, Not Larger Chemical Rockets
Since his undergraduate years, Yang has viewed fusion as something that “someone will do sooner or later, whether I do it or not.” On a 10-year horizon, he believes it could be the single development in physics with the greatest impact on civilization.
Once energy is extremely cheap, it will be difficult to predict the specific applications civilization creates, but demand will inevitably produce new ways to consume energy. Many products dependent on slow natural processes could shift toward energy-driven artificial synthesis.
Further out, he hopes fusion can enable propellantless space propulsion, freeing spacecraft from the propellant constraints of chemical rockets and creating genuinely commercial interplanetary travel. But he explicitly places that after “first making power generation work.”
38. The Organization Itself Must Be a Machine for Rapidly Creating New Experts
Because the market has no ready-made talent, Yang wants the team to turn people who are smart, ambitious, responsible and resilient—but unfamiliar with fusion engineering—into top specialists in a particular field within a short period.
The training path remains first-principles based: start from equations, run simulations, calibrate models through experiments and finally deliver physical equipment. Real capability is not getting the right answer by “flipping a coin and landing heads” once, but being able to systematically get it right again.
Hardware organizations differ from internet companies: roles cannot be divided too narrowly. Someone may work on one system today and need to learn another from scratch tomorrow if that team is short-staffed. A mistake involving a single screw, washer or nut can force the entire system into rework.
The team studies ITER’s Final Design Report and other design materials, as well as completed high-parameter machines and CFS/SPARC. The value is not merely copying parameters, but understanding the trade-offs made at the time and why manufacturing ultimately caused the result to deviate from the design.
39. This Startup Is Like Climbing a Mountain That Rises Exponentially but Is Believed to Have a Path Up
The team typically starts at 9 a.m. and finishes around 8 p.m.; overnight work is common during experiments. Team-building is rare, the probationary screening is strict, and people willing to join have “already prepared themselves to give up some parts of their lives.”
Yang acknowledges that negative feedback far outweighs positive encouragement and that he can be harsh in reprimanding people. He defines PUA as a manager harming someone for personal gain while packaging it as “for your own good.” The long-term core team and the company share the goal of making fusion work.
He sees 3 substantive reasons for company failure: not enough people, not enough money, or failure to get the work done. A lack of money means the project cannot start; securing resources but failing to deliver at the required cost and performance means no product can be formed. Scientifically, they have seen no no-go: “There are a pile of problems, and our judgment is that all of them have solutions.”
The closest analogy is not walking a tightrope, but continuously climbing a mountain whose height rises exponentially. Equipment and capability improve too, but the next section is always higher. “As long as you bring enough food, you will definitely make it up there—and you must not give up.” If you fall, “practice more,” then find another chance to keep climbing.